[PATCH v2 0/4] KVM: x86: Honor EFER_LMSLE_MBZ

[Jim Mattson <jmattson@google.com>]

v1 was a single patch that made KVM reject EFER.LMSLE=1 when the guest's
CPUID enumerates EFER_LMSLE_MBZ, CPUID.80000008H:EBX[bit 20].  Review turned
up two more things: KVM's *own* enumeration of the defeature is wrong today,
and the new guest-CPUID check needs an escape hatch so that userspace can
still set the bit on a host where KVM doesn't advertise it.

The motivation, as discussed in the v1 thread, is to be able to defeature a
virtual Rome, i.e. to set EFER_LMSLE_MBZ even though Rome itself supports
long mode segment limits, so that the vCPU can be hosted on Milan and later.
The opposite direction isn't interesting; in general you can't host
generation N+1 on generation N.

Patch 1 fixes KVM's enumeration of the defeature.  Today KVM passes
hardware's EFER_LMSLE_MBZ through as-is, i.e. leaves the bit clear on Intel
and on AMD with kvm_amd.nested=0, which tells userspace that long mode
segment limits *are* available.  But KVM allows EFER.LMSLE if and only if
nested SVM is supported, so on exactly those hosts KVM then rejects
WRMSR(EFER) with EFER.LMSLE=1.  Set EFER_LMSLE_MBZ whenever KVM refuses
EFER.LMSLE; the bit means precisely "EFER.LMSLE must be zero".  Note, this
is guest visible for userspace that reflects KVM's supported CPUID into the
guest.

Patch 2 is v1, plus the partial-emulation hunk Sean suggested so that
userspace can set EFER_LMSLE_MBZ on a host that does support LMSLE.  It
also masks EFER_LMSLE out of the value consumed by efer_trap().  Under
SEV-ES, EFER writes are *trapped*, not intercepted.  Rejecting the write
would inject a #GP *and* leave EFER.LMSLE set, which is strictly worse than
honoring a write that hardware allowed.

Patches 3 and 4 rename svm_nested_clear_efer_svme to svm_nested_efer_test
and add coverage for the defeature.

Tested on Rome, which supports LMSLE and so actually exercises the emulated
path, and on Skylake.

v1: https://lore.kernel.org/all/20260918154530.4129698-1-jmattson@google.com

Jim Mattson (4):
  KVM: x86: Advertise EFER_LMSLE_MBZ when KVM disallows EFER.LMSLE
  KVM: x86: Honor the guest's EFER_LMSLE_MBZ
  KVM: selftests: Rename svm_nested_clear_efer_svme to
    svm_nested_efer_test
  KVM: selftests: Add coverage for the EFER_LMSLE_MBZ defeature

 Documentation/virt/kvm/api.rst                |  25 ++
 arch/x86/kvm/cpuid.c                          |  20 ++
 arch/x86/kvm/msrs.c                           |  12 +-
 arch/x86/kvm/svm/svm.c                        |  21 +-
 arch/x86/kvm/vmx/vmx.c                        |   7 +
 arch/x86/kvm/x86.c                            |   9 +-
 tools/testing/selftests/kvm/Makefile.kvm      |   2 +-
 .../selftests/kvm/include/x86/processor.h     |   1 +
 .../kvm/x86/svm_nested_clear_efer_svme.c      |  50 ----
 .../selftests/kvm/x86/svm_nested_efer_test.c  | 273 ++++++++++++++++++
 10 files changed, 366 insertions(+), 54 deletions(-)
 delete mode 100644 tools/testing/selftests/kvm/x86/svm_nested_clear_efer_svme.c
 create mode 100644 tools/testing/selftests/kvm/x86/svm_nested_efer_test.c


base-commit: f0100363d8c374bd8e9ea7c9ba02744f0b802ca4
-- 
2.56.0.rc1.315.gc6ed9934b7-goog